The Khronos Group, the industry body that maintains OpenGL, OpenCL, EGL, glTF, Vulkan, OpenXR, and several other standards, has announced the Vulkan Safety Critical (SC) Working Group at Embedded World Conference 2019. The goal is to create an API that leverages Vulkan’s graphics and compute capabilities in a way that implementations can be safe and secure enough for the strictest of industries, such as automotive, air, medical, and energy.
It's a safety hammer, I promise. (No I don't.)
The primary goal is graphics and compute, although the working group will also consider exposing other hardware capabilities, such as video encode and decode. These industries currently have access to graphics through OpenGL SC, although the latest release is still significantly behind what a GPU can do. To put it into perspective – the latest OpenGL SC 2.0 (which was released in 2016) has less functionality than the original release of WebGL back in 2011.
While OpenGL SC 2.0 allows programmable vertex and fragment (pixel) shaders, it falls short in many areas. Most importantly, OpenGL SC 2.0 does not allow compute shaders; Vulkan SC is aiming to promote the GPU into a coprocessor for each of these important industries.
There is not much else to report on at this point – the working group has been formed. A bunch of industry members have voiced their excitement about the new API’s potential, such as Codeplay, Arm, and NVIDIA. The obvious example application would be self-driving cars, although I’m personally interested in the medical industry. Is there any sort of instrument that could do significantly more if it had access to a parallel compute device?
If you are in a safety-critical enterprise, then look into joining the Khronos Group.
“I’m personally interested in
“I’m personally interested in the medical industry. Is there any sort of instrument that could do significantly more if it had access to a parallel compute device?”
Medical scanners maybe(Only for analysing the output) but not the parts that controls any irradiation settings or actual irradiation dosing! I’d rather that stay with the minimum of software/hardware functionality that can be properly vetted/certified to be as error free as possible. So that’s sort of like having minimal software and triple redundancy where 3 Processors(Simple CPUs) have to all agree sort of things. And GPUs are already used for cat scans/CT Scans and MRI scans data analysis and visulation processing and the visual presentation.
GPUs for analysing the XRay data and digitally enhancing that data for visual presentation but out the loop on full control over and actual eguipment’s command and control functionality.
When it comes to massive amounts of shader cores controlling any safty critical processing I rather not have any safety critical compute/control done on any GPU shader cores! There are just to many sorts of complex interactions to properly test in less than 100+ year timelines for safty critical applications/direct GPU control over any plane/trains/Buses/automobiles and rockets xray/scanners and other such devices.
Most of those devices use very simple CPU cores that have only the bare minimal instruction sets that are designed for fault tolerance with triple+ redundancy and ECC everywhere on the processor. GPU processing power is best for analysing the medical imaging after the imaging data is collected and/or being streamed but are to complex to vett/certify for actually doing any of the irradiating and such where that process needs to be as error free as possible.
GPU’s can be great for data acquisition on self driving cars but that data must be presented to a very basic triple+ redundant CPU processor system for the actual decision making process, ditto for any other self driving vehicles/machines(Medical/others).
That’s why rockets and space probes are making use of those older specilized CPU processor designs with minimal instruction sets based on very simple hardware designs that can be easily vetted/certified in their hardware to be as close to error free as possible and still triple+ redundancy is made use of for any rocket parts where lives can be lost if things go wrong.
So sure if there is a SC-Vulkan API that assures that any GPU data analytics is as error free as possible from a GPU when that data is then presented by the GPU to the command and control triple+ redundant CPU processor based decision making process, then that’s a plausible usage. But GPU’s used for the actual command and control, that’s going to be very hard to properly vett/certify as it’s already damn hard to do that vetting and certification on those sinmple CPU cores that they make use of in rockets/space probes aircraft and other safty critical systems.
This(1) is what GPUs are good
This(1) is what GPUs are good for when they have Tensor Cores onboard and even when their shader sores are running the tensor flow libraries. But the triple redundant CPU(Simple CPU) cores that all have to agree on which way to steer the car or apply the brakes when that kid chasing after that ball comes out of nowhere into the car’s path.
And 90 seconds is too damn long for real time but is plenty of time for actually training the AI that’s then pre-loaded into the tensor cores after the learning is already done.
(1) [The sub-headline is a bit disingenuous about what that accuracy actually is about but read the article and you will see]
“This image-recognition neural net can be trained from 1.2 million pictures in the time it takes to make a cup o’ tea
Just 90 seconds, it’s claimed, provided a) you have 512 Nvidia V100 GPUs and b) er, no need for accuracy”
https://www.theregister.co.uk/2019/02/26/ai_training_world_record/
Edit: shader sores
to: shader
Edit: shader sores
to: shader cores
Though I’d imagine that I’d have some sores if I attempted to do all that number crunching using only pencil and paper! Me and a more than a few generations of others to get that much calculation work done that way!
Vulkan Safety Critical (SC)
Vulkan Safety Critical (SC) would like Vulkan hardened to be without any critical bugs? Or what am i missing here?
This would be like the LTS version has less of the latest and greatest features but is hardened to have less bugs?
Sounds like I would want some of the 2D platformers games to be made with this going forward…
Safety management